Skip to the result

ResearchTechnical note 12

Exact laws from data with wrong answers in it

Exact methods are usually fragile: one bad example and they fail. METRON learned the exact laws of machine arithmetic from data in which up to one answer in twenty was wrong.

Measured
9 November 2026
Wrong labels
1 % and 5 %
Laws
162
Record
§185

At a glance

162

laws learned exactly with 5 % of the training answers wrong.

1 in 20

examples carried a wrong answer, in training and in validation.

10

units complete again when the noise was drawn afresh.

In plain words

Imagine learning the multiplication table from a book with a misprint on every page. A careful student notices that the misprints do not fit the pattern and learns the real table anyway. METRON does the same, and the table it learns has no errors in it.

For engineers

Five 32-bit units (addition, comparison, word multiplexer, a rotation stage, a multiplication stage) learned with label-noise rates of 1 % and 5 % in both training and validation data, then verified for exactness on clean data. Repeated with an independent noise draw.

Why it matters

Field data is never clean. A method that is exact only in the laboratory has no market; one that stays exact on imperfect data can be pointed at real logs, real sensors and real traces.

MTRMETRON is AIDON’s exact co-processor. It learns laws from data, proves them and runs them, alongside a neural model or on its own.

The bottleneck

Exact methods break on the first bad example.

Classical exact learning assumes every example is correct. A single corrupted record makes the problem unsolvable, or worse, yields a confident wrong answer.

Statistical methods tolerate noise but give up exactness. Until now the choice was one or the other.

Result

Wrong answers in. Exact laws out.

Laws learned exactly, out of 162

Clean labels162 / 162
1 % of labels wrong162 / 162
5 % of labels wrong162 / 162

Fig. 1 Five units of machine arithmetic and logic, 162 laws in all, learned from data in which 1 in 100 and then 1 in 20 answers were deliberately wrong. Every law still came out exact.

Table 1 Learning from mislabelled data.
MeasureValue
Units learned (addition, comparison, selection, rotation, a multiplication step)5 of 5 complete
Laws learned exactly, clean labels162 of 162
Laws learned exactly, 1 % wrong labels162 of 162
Laws learned exactly, 5 % wrong labels162 of 162
Repeated with a fresh draw of noise10 of 10 units complete
Whole run7 min 15 s

Method

How METRON learns through noise

  1. Data1 % or 5 % of the answers are wrong
  2. METRONseparates the law from the errors
  3. Exact lawverified on clean data
  1. Wrong answers are injected on purpose, at a known rate, in the data METRON learns from and in the data it validates on.

  2. METRON weighs the evidence statistically and commits only to laws the data supports.

  3. Each law is then checked for exactness against the true function.

Controls

What was checked

Clean baseline

Without noise, the same 162 laws are learned, as in the earlier measurements.

Fresh noise

The whole test was repeated with a new, independent draw of wrong labels: 10 of 10 units complete.

Noise in validation too

The wrong answers are also in the validation data, so METRON has no clean reference to lean on.

Exactness on clean data

A law counts only if it is right on every output when checked against the truth.

Open questions

What this opens

  • Higher noise rates and noise that is not random.
  • Missing values and partially observed inputs.

Fields of application

Where it fits

  • Learning from field data: logs, sensors, traces.
  • Reverse engineering of hardware observed through an imperfect probe.
  • Any pipeline where the labels come from people or from another model.

Record

How this result is kept

Record
§185, measured
Run
9 November 2026, 7 min 15 s
Sample
5 units, 162 laws; 2 noise rates; 2 independent noise draws
Measurement file
SHA-256 5f68004cbc82a95f075de180b7bf228499bbecc74a02457eaf4788a9e6e1538e

Every figure on this page was read by a program from the raw measurement file, not typed. The fingerprint identifies that file: when it is disclosed, anyone can check that it is the same one.

Contact

Talk to the founders

Get in touch. We would be glad to share our ideas and walk you through the results, the proof and the measurements.